Scatter matrix based class separability criterion is commonly used in supervised feature extraction. But calculations of scatter matrixes depend on labeled data, so this criterion can not be used in unsupervised pattern. This paper presents a method to extend scatter matrix based class separability criterion to unsupervised pattern by fuzzy theory. The basic idea is to optimize the defined fuzzy Fisher criterion function to figure out fuzzy scatter matrixes in unsupervised pattern. Based on the obtained fuzzy between-class scatter matrix and fuzzy within-class scatter matrix, a novel class separability criterion based unsupervised feature extraction is proposed. Experimental results on its applications in UCI datasets show its effectiveness.